端到端微调整个 RAG 架构（含 DPR 检索器）用于问答
信息检索
2024-10-08 v1 人工智能
计算与语言
摘要
本文阐述了如何以端到端方式微调整个检索增强生成（Retrieval Augment Generation, RAG）架构。我们强调了实现该目标所需解决的主要工程挑战。我们还比较了端到端 RAG 架构在问答任务上如何优于原始 RAG 架构。我们已在 HuggingFace Transformers 库中开源了我们的实现。
关键词
引用
@article{arxiv.2106.11517,
title = {Fine-tune the Entire RAG Architecture (including DPR retriever) for Question-Answering},
author = {Shamane Siriwardhana and Rivindu Weerasekera and Elliott Wen and Suranga Nanayakkara},
journal= {arXiv preprint arXiv:2106.11517},
year = {2024}
}
备注
for associated code, see https://github.com/huggingface/transformers/tree/master/examples/research_projects/rag-end2end-retriever